Crossformer

Why we included this project

Crossformer comes from an ICLR 2023 paper, and this repo is the PyTorch code the authors released alongside it. The model treats each variable as its own dimension, groups nearby time points into segments, and runs attention both across time and across variables. A small set of router tokens keeps that cross-dimension stage linear in the number of series, which matters when you have many correlated variables to forecast at once. The repo ships training and evaluation scripts for the standard benchmarks (ETTh1, ETTm1, Weather, ECL, ILI, Traffic) plus a worked example for adapting the model to your own CSV data. If you're benchmarking transformer-based forecasters, it's a useful baseline and a good place to borrow architectural ideas from.